Center for Humane Technology

Stated Theory of Change

CHT claims a clear causal chain: Clarity about tech incentives -> Public/policymaker awareness -> Political pressure + litigation -> Changed incentives -> Better technology outcomes.

The mechanism has three prongs:

1. Mass media (documentaries, podcast, TED talks) to create widespread understanding that technology harms stem from misaligned business incentives, not individual bad actors
2. Elite engagement (Congressional testimony, Davos, policymaker briefings) to translate awareness into regulatory action
3. Litigation (expert consulting on lawsuits, product liability frameworks) to impose costs that change corporate behavior

Since 2023, this has been applied to AI: the same incentive dynamics that made social media harmful (race for engagement, ignoring externalities) are now driving an AI race that's even more dangerous. The proposed interventions are: product liability for AI harms, restrictions on AI companions for children, whistleblower protections, and what Harris calls the "narrow path" between chaotic decentralization and dystopian centralization.

Revealed Theory of Change

CHT's actual behavior reveals several things about what it optimizes for:

Attention over precision. CHT's media products consistently prioritize dramatic impact over epistemic accuracy. The Social Dilemma anthropomorphized algorithms as manipulative "bros" while experts warned against anthropomorphization. The AI Dilemma claimed "2024 will be the last human election" and presented the "50% extinction probability" statistic without methodological caveats. Neither claim was subsequently corrected. This pattern suggests CHT prioritizes moving the Overton window over getting the details right -- a defensible strategy for an advocacy organization, but one that creates credibility risk.

Harris's personal brand is the product. Harris is featured in three major documentaries, gives $100K-$200K+ speaking engagements, has appeared on every major media platform, and is the sole consistent leader across three Executive Directors. CHT's impact is almost entirely a function of Harris's personal credibility and media skills. The organization appears to be structured around amplifying Harris's voice rather than building institutional capacity.

Litigation is becoming the real lever. The Meta trial verdicts (2026) and Character.AI lawsuits represent CHT's most concrete impact mechanism. Aza Raskin's testimony -- drawing on his personal invention of infinite scroll -- provides unique credibility. Court-ordered injunctive relief (design changes) would be a direct realization of the "transforming incentives" theory of change. This is where the stated and revealed theories most closely align.

Not building bridges to technical safety community. Despite pivoting to AI existential risk in 2023, CHT has not engaged with the organizations doing technical alignment research. No collaborations with MIRI, ARC, Redwood, METR, or any technical safety org. No presence on EA Forum or LessWrong. No shared funders. This suggests CHT views its value-add as entirely in the mainstream communication space and has made a deliberate choice not to engage with the technical community.

Key Assumptions

Assumption 1: Public awareness leads to policy change.

• Evidence for: The Social Dilemma preceded a wave of tech regulation proposals. 47 states passed child safety laws in the years following. Meta/Google trial verdicts in 2026 demonstrate accountability.
• Evidence against: Causal attribution is extremely weak. These regulatory trends have many drivers. Facebook's fundamental business model has not changed despite massive public awareness of its harms. The attention economy is more dominant in 2026 than in 2018.
• Testable? Partially -- you can track whether specific CHT-associated policy proposals (AI LEAD Act, Character.AI injunctive relief) are enacted.
• If wrong: CHT is generating noise without signal. Awareness without institutional change is merely entertainment.

Assumption 2: AI risk can be communicated to mainstream audiences without dangerous oversimplification.

• Evidence for: CHT reaches audiences that technical AI safety researchers simply cannot. 100M viewers is not something any amount of alignment research papers can achieve.
• Evidence against: The AI Dilemma contains significant oversimplifications and some claims that have aged poorly. Nir Eyal's critique suggests the "powerless voodoo doll" framing may actually be counterproductive by inducing learned helplessness. LibrarianShipwreck argues the simplified framing protects tech companies by shifting blame to abstract "incentive structures" rather than specific actors.
• If wrong: CHT is creating a public understanding of AI risk that is so distorted it leads to wrong policy responses (e.g., regulations that entrench incumbents rather than reduce risk).

Assumption 3: CHT's independence from AI labs is a feature, not a bug.

• Evidence for: No financial ties to any AI lab means no incentive to pull punches. CHT can criticize Anthropic, OpenAI, and Meta without worrying about funding cuts.
• Evidence against: Independence from the technical community means CHT may misrepresent the state of AI safety research. No technical staff means no ability to evaluate AI claims independently.
• If wrong: CHT is making policy recommendations in a domain it doesn't technically understand, potentially leading to harmful regulation.

Assumption 4: The "insider critic" model is credible rather than captured.

• Evidence for: Harris and Raskin have specific technical knowledge (Google design ethics, infinite scroll invention) that gives their testimony weight. Raskin's Meta trial testimony was uniquely powerful because he invented the mechanism being criticized.
• Evidence against: LibrarianShipwreck's core critique is that insiders-as-critics serve to control the discourse and protect the industry. Harris repeatedly telling Congress "they're not evil" provides exactly the cover tech companies need. The critique is not about intent but about structural position.
• If wrong: CHT's criticism is the acceptable face of dissent that lets the system continue unreformed.

Strengths

Unmatched mainstream reach. No other organization in the AI safety ecosystem comes close to CHT's public reach. 100M documentary viewers, 27.5M podcast downloads, TED talks, Davos presence, Congressional testimony. CHT reaches people who will never read an alignment research paper or visit LessWrong.

Financial independence. No AI lab funding, no Coefficient Giving/Open Phil money, no EA ecosystem ties. This gives CHT a genuinely independent voice that few AI safety organizations have.

Concrete litigation outcomes. The Meta trial verdicts and pending injunctive relief represent the most direct realization of the "change incentives" theory of change. Court-ordered product design changes would be unprecedented.

Positioned for bipartisan credibility. CHT works with both Democratic and Republican senators. In an era of increasing politicization of tech regulation, this bipartisan positioning is valuable and rare.

Raskin's unique credibility. Having invented infinite scroll and then testifying about its harms gives Raskin a credibility that no external critic could achieve. This is a genuine, non-replicable asset.

Weaknesses and Risks

No technical depth on AI. CHT pivoted to AI existential risk without hiring AI researchers, building technical capacity, or engaging with the technical alignment community. The AI Dilemma contains oversimplifications that technical researchers would flag. This creates a risk that CHT's AI advocacy is well-intentioned but technically uninformed.

Key-person dependency. Harris is the organization. Three EDs have come and gone while Harris remains the constant. If Harris's credibility is damaged or he departs, CHT has no institutional buffer.

Epistemic carelessness. Specific claims in the AI Dilemma have aged poorly ("2024 will be the last human election") without correction. The "50% of AI researchers" statistic is presented without survey methodology. The Social Dilemma was critiqued for taking Harris quotes out of context. An organization devoted to "creating clarity" should model rigorous self-correction.

Financial precarity. Revenue swung from $9.5M to $3.6M in one year while expenses nearly doubled. At FY2024 burn rate, CHT has roughly one year of runway. The organization's financial health depends on continued foundation support in an environment where AI is becoming politically fraught.

Speaking fee incentives. Harris reportedly earns $100K-$200K+ per speaking engagement, potentially exceeding his CHT salary. This creates a personal financial incentive toward alarm-level rhetoric that commands premium speaker fees, whether or not it perfectly tracks with the evidence.

The "insider critic" trap. The LibrarianShipwreck critique is difficult to dismiss: by maintaining that tech companies "mean well" and that the problem is incentive structures rather than corporate malice, CHT may provide the industry with its most valuable form of cover -- criticism that ultimately reinforces the system it claims to challenge.

Cross-References

Complementary to technical AI safety orgs (MIRI, ARC, Redwood, METR): CHT reaches audiences these orgs never will. The technical community produces the research; CHT translates risk awareness to the public and policymakers. The danger is that CHT's translation is inaccurate.

Overlapping with but distinct from PauseAI and ControlAI: These orgs also push for AI policy intervention, but from within the EA/rationalist ecosystem. CHT operates independently with mainstream foundation funding and mainstream media access. There is no formal coordination.

In tension with frontier labs' safety teams: CHT criticizes the race dynamics that labs claim to be managing internally. Anthropic's endorsement of product liability (noted by Harris) suggests some lab safety teams may view CHT as a useful external pressure mechanism.

Different from CAIS/FLI: These orgs also bridge policy and safety but have deeper technical engagement. FLI gave CHT $500K in 2023, suggesting it views CHT's mainstream reach as valuable.

Parallel to Jonathan Haidt's work: Haidt (The Anxious Generation) focuses on similar social media harms with a more research-grounded approach. He has appeared on CHT's podcast. Their work is mutually reinforcing.

What Would Change This Assessment

• CHT hires technical AI safety staff and begins engaging with alignment research: Would significantly increase credibility of AI risk claims.
• Injunctive relief from Meta trial mandates specific design changes: Would be the strongest validation of CHT's theory of change ever produced.
• Harris is caught making knowingly false or misleading claims: Would devastate CHT's credibility, which is entirely based on Harris's perceived authenticity.
• Technical AI safety community publishes substantive critique of CHT's AI framing: Would clarify whether CHT's mainstream communication of AI risk is helpful or harmful.
• CHT's FY2025 financials show continued revenue decline: Would suggest the organization is not financially sustainable and may be forced to compromise its independence.
• Published evidence linking CHT's awareness work to specific policy enactment (not just correlation): Would resolve the fundamental question of whether awareness campaigns actually change incentives.

Self-Critique

Weakest claim: My assessment that CHT's litigation work represents a genuine validation of its theory of change may be premature. The Meta trial verdicts are pending appeal, and injunctive relief has not yet been specified. I may be giving too much weight to early-stage legal outcomes.

Potential bias: I may be underweighting the LibrarianShipwreck critique because it is rhetorically aggressive and from an academic rather than technical perspective. The core argument -- that insider critics serve the system they critique -- is a structural argument that is very difficult to evaluate empirically. I have treated it as one perspective among many rather than as the primary lens.

What I should have checked: I was unable to access the EA Forum linkpost about the AI Dilemma, which might contain the only substantive engagement between the technical AI safety community and CHT's public messaging. I also could not access the full Chuck Russell Medium article, which is reportedly the most detailed recent critical examination of Harris.

What a thoughtful disagreer would say: "CHT reaches 100M people with an imperfect message about AI risk. The alternative is that those 100M people hear nothing about AI risk at all. Perfect is the enemy of good. The technical community's failure to communicate with mainstream audiences is not CHT's fault, and criticizing CHT for oversimplification is easy from the sidelines."

Single weakest element: The absence of any published assessment from technical AI safety researchers about whether CHT's mainstream AI risk communication is accurate, helpful, or harmful. Without this, I am forced to evaluate CHT's AI claims using my own judgment rather than expert evaluation, which is less reliable.